Nombre: MATHEUS VIEIRA LESSA RIBEIRO
Fecha de publicación: 11/10/2017
Supervisor:
Nombre |
Papel |
|---|---|
| EVANDRO OTTONI TEATINI SALLES | Advisor * |
Junta de examinadores:
| Nombre |
Papel |
|---|---|
| AURA CONCI | External Examiner * |
| EVANDRO OTTONI TEATINI SALLES | Advisor * |
| KELLY ASSIS DE SOUZA GAZOLLI | External Examiner * |
Sumario: This master thesis proposes a new visual descriptor of scenes from Local Binary Pattern (LBP) and exploring a spatial information utilizing the Color Coherent Vector (CCV) algorithm. LBP is characterized by a non-linear and non-parametric technique, it does not use intermediate concepts in the process of image description, becoming an alternative for lay users with low knowledge in this area. In contrast, CCV has proved to be a safe technique in minimizing the problem of lack of information by histograms, it expresses an image by
coherent and incoherent pixels with no needs to improve the dimensionality of data. In this way, a first approach was the proposal of LBP Incoherent e LBP Coherent techniques in the scene classification. Preliminary outcomes, with K-NN classifier, indicated that LBP Incoherent performs a good compromise between accuracy and dimension of data representation. Afterwards, with the purpose to including the concept of context, to minimizing the problem of location from LBP, the Contextual Modified Local Binary Pattern Incoherent
(CMLBP Incoherent) was proposed, which models the distributions of local structures through LBP, by adding contextual information, inspired in Contextual Modified Census Transform (CMCT). The CMLBP Incoherent, among others characteristics, has demonstrated competency in discarding homogeneous regions, represented by coherent pixels, through CCV algorithm.
In experiments carried with important datasets in literature, CMLBP achieved better results than original techniques which do not discard the coherent pixels, in almost all over tests. For scenes with much detail and information the results were satisfactory and better than the results of known techniques in the literature. The results obtained by CMLBP Incoherent has been encouraging to finding of a descriptor with good discriminant performance and low dimensionality in image representation.
